def main(_):
# 1. change the dataset
# dataset = imagenet.get_split('train'), FLAGS.data_dir)
dataset = flowers.get_split(FLAGS.data_split, FLAGS.data_dir)
model = InceptionModel(checkpoints_file=FLAGS.checkpoint_file_path)
# 2. set the model to training mode
# op, graph model.build(dataset, image_height=224, image_width=224, num_classes=1000, is_training=True)
op, graph = model.build(dataset,
image_height=224,
image_width=224,
num_classes=1000,
is_training=False)
# 3. comment out the actual training code
# slim.learning.train(
# op,
# logdir=train_dir,
# init_fn=model.init_fn,
# number_of_steps=100)
# 4. dump model to the specified path
from bigdl.util.tf_utils import dump_model
dump_model(path=FLAGS.dump_model_path,
ckpt_file=FLAGS.checkpoint_file_path,
graph=graph)
def main():
"""
How to run this script:
python export_tf_checkpoint.py meta_file chkp_file save_path
"""
saver = tf.train.import_meta_graph(argv[1])
with tf.Session() as sess:
saver.restore(sess, argv[2])
dump_model(argv[3], sess)
def main():
meta_file = None
checkpoint = None
save_path = "model"
saver_folder = None
if len(argv) == 2:
if op.isdir(argv[1]):
saver_folder = argv[1]
else:
meta_file = argv[1] + ".meta"
checkpoint = argv[1]
elif len(argv) == 3:
if op.isdir(argv[1]):
saver_folder = argv[1]
else:
meta_file = argv[1] + ".meta"
checkpoint = argv[1]
save_path = argv[2]
elif len(argv) == 4:
meta_file = argv[1]
checkpoint = argv[2]
save_path = argv[3]
else:
print(
"Invalid script arguments. How to run the script:\n" +
"python export_tf_checkpoint.py checkpoint_name\n" +
"python export_tf_checkpoint.py saver_folder\n" +
"python export_tf_checkpoint.py checkpoint_name save_path\n" +
"python export_tf_checkpoint.py saver_folder save_path\n" +
"python export_tf_checkpoint.py meta_file checkpoint_name save_path"
)
exit(1)
if op.isfile(save_path):
print("The save folder is a file. Exit")
exit(1)
if not op.exists(save_path):
print("create folder " + save_path)
os.makedirs(save_path)
with tf.Session() as sess:
if saver_folder is None:
saver = tf.train.import_meta_graph(meta_file, clear_devices=True)
saver.restore(sess, checkpoint)
else:
tf.saved_model.loader.load(sess,
[tf.saved_model.tag_constants.SERVING],
saver_folder)
checkpoint = save_path + '/model.ckpt'
saver = tf.train.Saver()
saver.save(sess, checkpoint)
dump_model(save_path, None, sess, checkpoint)
def main():
meta_file = None
checkpoint = None
save_path = "model"
saver_folder = None
if len(argv) == 2:
if op.isdir(argv[1]):
saver_folder = argv[1]
else:
meta_file = argv[1] + ".meta"
checkpoint = argv[1]
elif len(argv) == 3:
if op.isdir(argv[1]):
saver_folder = argv[1]
else:
meta_file = argv[1] + ".meta"
checkpoint = argv[1]
save_path = argv[2]
elif len(argv) == 4:
meta_file = argv[1]
checkpoint = argv[2]
save_path = argv[3]
else:
print("Invalid script arguments. How to run the script:\n" +
"python export_tf_checkpoint.py checkpoint_name\n" +
"python export_tf_checkpoint.py saver_folder\n" +
"python export_tf_checkpoint.py checkpoint_name save_path\n" +
"python export_tf_checkpoint.py saver_folder save_path\n" +
"python export_tf_checkpoint.py meta_file checkpoint_name save_path")
exit(1)
if op.isfile(save_path):
print("The save folder is a file. Exit")
exit(1)
if not op.exists(save_path):
print("create folder " + save_path)
os.makedirs(save_path)
with tf.Session() as sess:
if saver_folder is None:
saver = tf.train.import_meta_graph(meta_file, clear_devices=True)
saver.restore(sess, checkpoint)
else:
tf.saved_model.loader.load(sess, [tf.saved_model.tag_constants.SERVING], saver_folder)
checkpoint = save_path + '/model.ckpt'
saver = tf.train.Saver()
saver.save(sess, checkpoint)
dump_model(save_path, None, sess, checkpoint)
def main(_):
# 1. change the dataset
# dataset = imagenet.get_split('train'), FLAGS.data_dir)
dataset = flowers.get_split(FLAGS.data_split, FLAGS.data_dir)
model = InceptionModel(checkpoints_file=FLAGS.checkpoint_file_path)
# 2. set the model to training mode
# op, graph model.build(dataset, image_height=224, image_width=224, num_classes=1000, is_training=True)
op, graph = model.build(dataset, image_height=224, image_width=224, num_classes=1000, is_training=False)
# 3. comment out the actual training code
# slim.learning.train(
# op,
# logdir=train_dir,
# init_fn=model.init_fn,
# number_of_steps=100)
# 4. dump model to the specified path
from bigdl.util.tf_utils import dump_model
dump_model(path=FLAGS.dump_model_path, ckpt_file=FLAGS.checkpoint_file_path, graph=graph)
import tensorflow as tf # This is your model definition. xs = tf.placeholder(tf.float32, [None, 1]) W1 = tf.Variable(tf.zeros([1,10])+0.2) b1 = tf.Variable(tf.zeros([10])+0.1) Wx_plus_b1 = tf.nn.bias_add(tf.matmul(xs,W1), b1) output = tf.nn.tanh(Wx_plus_b1, name="output") # Adding the following lines right after your model definition from bigdl.util.tf_utils import dump_model dump_model_path = "/tmp/model" # This line of code will create a Session and initialized all the Variable and # save the model definition and variable to dump_model_path as BigDL readable format. dump_model(path=dump_model_path)